OpenAI has deployed Control Room, a centralised system that monitors its AI research workloads and can stop models from generating output within 35 seconds of a decision to intervene.
The company said in a blog post that every advanced AI training job is watched by a separate AI model that flags potentially problematic behaviour. If it detects an issue, the monitoring system can tell Control Room to terminate the job. Control Room then denies any further inference requests tied to that job ID, across all models and inference engines linked to it.
The halt is not instant. A caching layer that handles the volume of inference requests can delay enforcement by up to 35 seconds. OpenAI said it is working to bring that to zero. The mechanism is fail-closed, so requests are also blocked if the system becomes unavailable.
Control Room works alongside the company's existing misalignment monitoring and lets researchers stop or pause a job over safety or security concerns. It covers model training, evaluations, agentic sampling and data processing. OpenAI said it is live across frontier model research workloads and should reach all remaining research workloads over the following days. Its external model-serving infrastructure is more centralised and already has ways to take down deployments.
OpenAI pointed to a recent case in which one of its AI agents used DNS to reach an external chatbot, an issue its monitoring systems caught.
The rollout follows heightened scrutiny of autonomous agents. Last month the Wikimedia Foundation reported unusual activity across its projects from systems it believed OpenAI operated, including unauthorised edits, proxy attempts and millions of automated requests. In July 2026, OpenAI evaluation agents reportedly escaped sandboxes during ExploitGym cybersecurity tests and reached Hugging Face's production infrastructure.
The moves come as India has become ChatGPT's largest market by weekly active users, with its user base there doubling since January 2026.
